Carry out nearest neighbor search using JSON through a web-server application, written in Rustlang in BC, Canada.
Great for searching natural language vectors, image vectors, or any other vectors where you want the smallest distance between vectors. It accommodates geo-coordinates which is great for services or goods that a bound to a physical location, whether it's a store or a venue. It only returns an array of ID, vector distance and geographic distance in KMs, so if there's more data than just the ID, that needs to be reconciled with the rest of the data.
Required file format: unexited rows of byte data: 32 bytes for the ID, then some number of f32 for the vector (512 is just the example), then 2 of f32 for the coordinate.
To start: cargo run --release ./words_subset.data 8989 512
Where: cargo run --release path_to_data port_to_host vector_size
Request format is JSON to /search or /search_ann:
{"sort_by_vec": 1 or 0,
"geoc": [float],
"vector": [float], (matching the database size)
"geo_threshold": float,
"vec_threshold": float,
"limit_results": int}
Response format is JSON:
{"items: [{"id": str, "geo_dist": float, "dist": float}, ...]}
Demo.py should yield results on 10k records in this output using the words_subset.data file:
times for exhaustive search
max time taken 0:00:00.023136
min time taken 0:00:00.013079
median time taken 0:00:00.013513
times for approximate search
max time taken 0:00:00.001968
min time taken 0:00:00.001162
median time taken 0:00:00.001391
There are commented-out lines related to geo-distance in kilometers. Please feel free to reach out to me if you need some help getting that working. My email is andrew.matte@gmail.comst
I am using the kmeans crate to cluster the vectors so that I can search the cluster centroids in the /search_ann index. That crate relies on packed_simd_2 which I am able to build locally but not on some other machines, which makes the build fail on github's workflows. I am considering changing that package out for another so that more people can use it.
The search runs on a single core, but if multiple queries are executing at the same time, it can use more cores thanks to Actix, the webserver.
Made Andrew Matte, with some help from Bob Matte, and a secret third person. Mostly funded by the Canadian Government's Employment and Social Development Canada's (ESDC's) Youth Employment and Skills Strategy (YESS) through the Contextual Opportunities and Resources Distribution System (CORDS) project.
Over the past year I developed my own nearest neighbour search application for vectors because I've been disappointed with what's out there. I used Rust because it's my favourite compiled language out of the half-dozen I've tried. I prototyped the project in python to understand what had to be done, end to end, which took me about a day. Another tie to python is because most ML models are written in python so the input has to come from there. I kept finding improvements until I was tired of hunting for bottlenecks. When I moved to Rust, I started with what would work quickly: a console app that could accept an input, read the existing vectors, and stream the output of all of the distances. Once I was able to read all the inputs and compute the output without persistent memory, I decided to put it behind actix, a rust-based web server, so that it I could avoid loading times. This was not only an optimization to avoid re-reading the same data but also to accommodate JSON easier. Once that was done, I added automated indexing to the project, which I did with the Kmeans library in rust. There are still things I would do to the project, such as allowing the database to create, update, and delete records, and to do the same to the index. I want to switch out the clustering that I'm using for the index to a more supported library too.